gather the gradients in one vector

Print a memory summary.

No grads wrapper.

Learning forward. L2-embedding normalization.

Init prototoypes of new classes.

This stores parameter gradients of past tasks. pp: parameters grads: gradients grad_dims: list with number of paramet

Evaluates performance of the model on samples from all the tasks and reports 1) average performance on all the samples regardless of their ta

This stores the gradient of a new memory and compute the dot product with the previously stored memories. pp: parameters mem

Compute accuracy on the buffer.

This is used to overwrite the gradients with a new gradient vector, whenever violations occur. pp: parameters newgrad

Solves the GEM dual QP described in the paper given a proposed gradient "gradient", and a memory of task gradients "memories".

- \sum_{i in B^c} log(Pc) - \sum_{i in B^c} \sum_{k \ne c} log( (1 - Pk)) Note: log(Exp(y)) makes y always positive, which

update age of all examplars, and update representations of replayed exemplars.

Update to maintain

Class-specific memories are updated following defined strategy.

Standard reduction is mean, as we use full batch information instead of per-sample. Symmetry in the distance function inhibits summin

:param margin: margin on distance between pos vs neg samples (see TripletMarginLoss) :param dist: distance function 2 vectors (e.g. L

:param p_mode: prototypes update mode

:param sample_qi_mode: input memory queue update mode

:param qi_mode: input memory queue update mode

Compute offsets for cifar to determine which outputs to select for a given task.

Deployment forward. Find closest prototype for each sample.

Train.